一种改进的YOLOv8s模型用于电子线路板瑕疵检测.pdfVIP

长，信息通信

DOI:10.20153/j.issn.2096-9759.2025.01.016

一种改进的YOLOv8s模型用于电子线路板瑕疵检测

黄文浩

三峡大学计算机与信息学院，湖北宜昌443002

AnImprovedYOLOv8sModelforDefectDetectioninPrintedCircuitBoards

HUANGWenhao

ChinaThreeGorgesUniversity,CollegeofComputerandInfomationTechnology,Yichang443002

摘要：电子线路板在生产制作过程中存Abstract:PrintedCircuitBoardshavedefects,leakageholes,etc.,intheproductionandfabrica-

在瑕疵、缺陷、漏孔等问题，因此在生成tionprocess,Itisthereforeimportantandnecessarytodetectdefectsatthisstage,whichcan

制造过程中对其进行瑕疵检测是重要且quicklylocatetheproblemandmakeupforcertaineconomiclosses.Forthetaskofprintedcir-

必要的，这能快速定位问题与弥补一定cuitboarddefectdetection,thispaperproposesafine-grainedobjectdetectorcalledEFF-

的经济损失。针对电子线路板瑕疵检测YOLOv8sbasedonYOLOv8s.firstly,inordertoextractimagefeaturesatafinc-grainedlevel

任务，该文提出一种改进YOLOv8s模型andtoimprovetheobjectdetectionaccuracyoftheimage,EfficientFormerV2isutilizedasthe

的细粒度图像目标检测方法EFF-backbone,replacingtheoriginalconvolutionblockstructureoftheYOLOv8s.Secondly,inorder

YOLOv8s。首先，为了细粒度地提取图tomoreprominentlycharacterizetheinformationoftheobjcct,thispaperproposesanewC2f-

像特征与提高图像的目标检测准确率，ELAmodulebasedontheefficientspatialattentionduringthefeaturefusionstage.Thefinal

使用EfficientFormerV2作为主干网络，experimentalresultsshowthatTheaccuracy,recall,mAP@0.5andmAP@0.5:0.95ofthe

替换YOLOv8s模型原始的卷积块结构；methodproposedinthispaperontheprintedcircuitboarddefectdetectiondatasetcanreach

其次为了更突出地表征待检测目标的信97.3%、93.9%、97.1%和55.4%,respectively,andrise.2.7and4.6points,respec-

息，该文在特征融合阶段基于高效的空tively,comparedwiththeYOLOv8s.

间注意力ELA提出一种全新的C2f-ELAKeyWor